2 * Copyright (c) 2002-2006 Rice University
3 * Copyright (c) 2007-2008 Alan L. Cox <alc@cs.rice.edu>
6 * This software was developed for the FreeBSD Project by Alan L. Cox,
7 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
24 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
25 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
26 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
28 * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
33 * Superpage reservation management module
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
41 #include <sys/param.h>
42 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
46 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
52 #include <vm/vm_param.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_phys.h>
56 #include <vm/vm_reserv.h>
59 * The reservation system supports the speculative allocation of large physical
60 * pages ("superpages"). Speculative allocation enables the fully-automatic
61 * utilization of superpages by the virtual memory system. In other words, no
62 * programmatic directives are required to use superpages.
65 #if VM_NRESERVLEVEL > 0
68 * The number of small pages that are contained in a level 0 reservation
70 #define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
73 * The number of bits by which a physical address is shifted to obtain the
76 #define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
79 * The size of a level 0 reservation in bytes
81 #define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
84 * Computes the index of the small page underlying the given (object, pindex)
85 * within the reservation's array of small pages.
87 #define VM_RESERV_INDEX(object, pindex) \
88 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
91 * The reservation structure
93 * A reservation structure is constructed whenever a large physical page is
94 * speculatively allocated to an object. The reservation provides the small
95 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
96 * within that object. The reservation's "popcnt" tracks the number of these
97 * small physical pages that are in use at any given time. When and if the
98 * reservation is not fully utilized, it appears in the queue of partially-
99 * populated reservations. The reservation always appears on the containing
100 * object's list of reservations.
102 * A partially-populated reservation can be broken and reclaimed at any time.
105 TAILQ_ENTRY(vm_reserv) partpopq;
106 LIST_ENTRY(vm_reserv) objq;
107 vm_object_t object; /* containing object */
108 vm_pindex_t pindex; /* offset within object */
109 vm_page_t pages; /* first page of a superpage */
110 int popcnt; /* # of pages in use */
115 * The reservation array
117 * This array is analoguous in function to vm_page_array. It differs in the
118 * respect that it may contain a greater number of useful reservation
119 * structures than there are (physical) superpages. These "invalid"
120 * reservation structures exist to trade-off space for time in the
121 * implementation of vm_reserv_from_page(). Invalid reservation structures are
122 * distinguishable from "valid" reservation structures by inspecting the
123 * reservation's "pages" field. Invalid reservation structures have a NULL
126 * vm_reserv_from_page() maps a small (physical) page to an element of this
127 * array by computing a physical reservation number from the page's physical
128 * address. The physical reservation number is used as the array index.
130 * An "active" reservation is a valid reservation structure that has a non-NULL
131 * "object" field and a non-zero "popcnt" field. In other words, every active
132 * reservation belongs to a particular object. Moreover, every active
133 * reservation has an entry in the containing object's list of reservations.
135 static vm_reserv_t vm_reserv_array;
138 * The partially-populated reservation queue
140 * This queue enables the fast recovery of an unused cached or free small page
141 * from a partially-populated reservation. The reservation at the head of
142 * this queue is the least-recently-changed, partially-populated reservation.
144 * Access to this queue is synchronized by the free page queue lock.
146 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
147 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
149 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
151 static long vm_reserv_broken;
152 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
153 &vm_reserv_broken, 0, "Cumulative number of broken reservations");
155 static long vm_reserv_freed;
156 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
157 &vm_reserv_freed, 0, "Cumulative number of freed reservations");
159 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
161 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
162 sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
164 static long vm_reserv_reclaimed;
165 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
166 &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
168 static void vm_reserv_depopulate(vm_reserv_t rv);
169 static vm_reserv_t vm_reserv_from_page(vm_page_t m);
170 static boolean_t vm_reserv_has_pindex(vm_reserv_t rv,
172 static void vm_reserv_populate(vm_reserv_t rv);
173 static void vm_reserv_reclaim(vm_reserv_t rv);
176 * Describes the current state of the partially-populated reservation queue.
179 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
184 const int cbufsize = (VM_NRESERVLEVEL + 1) * 81;
185 int counter, error, level, unused_pages;
187 cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO);
188 sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN);
189 sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n");
190 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
193 mtx_lock(&vm_page_queue_free_mtx);
194 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
196 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
198 mtx_unlock(&vm_page_queue_free_mtx);
199 sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
200 unused_pages * ((int)PAGE_SIZE / 1024), counter);
203 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
210 * Reduces the given reservation's population count. If the population count
211 * becomes zero, the reservation is destroyed. Additionally, moves the
212 * reservation to the tail of the partially-populated reservations queue if the
213 * population count is non-zero.
215 * The free page queue lock must be held.
218 vm_reserv_depopulate(vm_reserv_t rv)
221 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
222 KASSERT(rv->object != NULL,
223 ("vm_reserv_depopulate: reserv %p is free", rv));
224 KASSERT(rv->popcnt > 0,
225 ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
226 if (rv->inpartpopq) {
227 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
228 rv->inpartpopq = FALSE;
231 if (rv->popcnt == 0) {
232 LIST_REMOVE(rv, objq);
234 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
237 rv->inpartpopq = TRUE;
238 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
243 * Returns the reservation to which the given page might belong.
245 static __inline vm_reserv_t
246 vm_reserv_from_page(vm_page_t m)
249 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
253 * Returns TRUE if the given reservation contains the given page index and
256 static __inline boolean_t
257 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
260 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
264 * Increases the given reservation's population count. Moves the reservation
265 * to the tail of the partially-populated reservation queue.
267 * The free page queue must be locked.
270 vm_reserv_populate(vm_reserv_t rv)
273 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
274 KASSERT(rv->object != NULL,
275 ("vm_reserv_populate: reserv %p is free", rv));
276 KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
277 ("vm_reserv_populate: reserv %p is already full", rv));
278 if (rv->inpartpopq) {
279 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
280 rv->inpartpopq = FALSE;
283 if (rv->popcnt < VM_LEVEL_0_NPAGES) {
284 rv->inpartpopq = TRUE;
285 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
290 * Allocates a page from an existing or newly-created reservation.
292 * The object and free page queue must be locked.
295 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex)
297 vm_page_t m, mpred, msucc;
298 vm_pindex_t first, leftcap, rightcap;
301 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
304 * Is a reservation fundamentally not possible?
306 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
307 if (pindex < VM_RESERV_INDEX(object, pindex) ||
308 pindex >= object->size)
312 * Look for an existing reservation.
315 mpred = object->root;
316 while (mpred != NULL) {
317 KASSERT(mpred->pindex != pindex,
318 ("vm_reserv_alloc_page: pindex already allocated"));
319 rv = vm_reserv_from_page(mpred);
320 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) {
321 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
322 /* Handle vm_page_rename(m, new_object, ...). */
323 if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
325 vm_reserv_populate(rv);
327 } else if (mpred->pindex < pindex) {
329 (msucc = TAILQ_NEXT(mpred, listq)) == NULL)
331 KASSERT(msucc->pindex != pindex,
332 ("vm_reserv_alloc_page: pindex already allocated"));
333 rv = vm_reserv_from_page(msucc);
334 if (rv->object == object &&
335 vm_reserv_has_pindex(rv, pindex)) {
336 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
337 /* Handle vm_page_rename(m, new_object, ...). */
338 if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
340 vm_reserv_populate(rv);
342 } else if (pindex < msucc->pindex)
344 } else if (msucc == NULL) {
346 mpred = TAILQ_PREV(msucc, pglist, listq);
350 mpred = object->root = vm_page_splay(pindex, object->root);
354 * Determine the first index to the left that can be used.
358 else if ((rv = vm_reserv_from_page(mpred))->object != object)
359 leftcap = mpred->pindex + 1;
361 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
364 * Determine the first index to the right that cannot be used.
367 rightcap = pindex + VM_LEVEL_0_NPAGES;
368 else if ((rv = vm_reserv_from_page(msucc))->object != object)
369 rightcap = msucc->pindex;
371 rightcap = rv->pindex;
374 * Determine if a reservation fits between the first index to
375 * the left that can be used and the first index to the right
376 * that cannot be used.
378 first = pindex - VM_RESERV_INDEX(object, pindex);
379 if (first < leftcap || first + VM_LEVEL_0_NPAGES > rightcap)
383 * Would a new reservation extend past the end of the given object?
385 if (object->size < first + VM_LEVEL_0_NPAGES) {
387 * Don't allocate a new reservation if the object is a vnode or
388 * backed by another object that is a vnode.
390 if (object->type == OBJT_VNODE ||
391 (object->backing_object != NULL &&
392 object->backing_object->type == OBJT_VNODE))
394 /* Speculate that the object may grow. */
398 * Allocate a new reservation.
400 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
402 rv = vm_reserv_from_page(m);
403 KASSERT(rv->pages == m,
404 ("vm_reserv_alloc_page: reserv %p's pages is corrupted",
406 KASSERT(rv->object == NULL,
407 ("vm_reserv_alloc_page: reserv %p isn't free", rv));
408 LIST_INSERT_HEAD(&object->rvq, rv, objq);
411 KASSERT(rv->popcnt == 0,
412 ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted",
414 KASSERT(!rv->inpartpopq,
415 ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE",
417 vm_reserv_populate(rv);
418 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
424 * Breaks all reservations belonging to the given object.
427 vm_reserv_break_all(vm_object_t object)
432 mtx_lock(&vm_page_queue_free_mtx);
433 while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
434 KASSERT(rv->object == object,
435 ("vm_reserv_break_all: reserv %p is corrupted", rv));
436 if (rv->inpartpopq) {
437 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
438 rv->inpartpopq = FALSE;
440 LIST_REMOVE(rv, objq);
442 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
443 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
444 vm_phys_free_pages(&rv->pages[i], 0);
448 KASSERT(rv->popcnt == 0,
449 ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
453 mtx_unlock(&vm_page_queue_free_mtx);
457 * Frees the given page if it belongs to a reservation. Returns TRUE if the
458 * page is freed and FALSE otherwise.
460 * The free page queue lock must be held.
463 vm_reserv_free_page(vm_page_t m)
467 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
468 rv = vm_reserv_from_page(m);
469 if (rv->object == NULL)
471 if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
472 vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
474 vm_reserv_depopulate(rv);
479 * Initializes the reservation management system. Specifically, initializes
480 * the reservation array.
482 * Requires that vm_page_array and first_page are initialized!
491 * Initialize the reservation array. Specifically, initialize the
492 * "pages" field for every element that has an underlying superpage.
494 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
495 paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
496 while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
497 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
498 PHYS_TO_VM_PAGE(paddr);
499 paddr += VM_LEVEL_0_SIZE;
505 * Returns a reservation level if the given page belongs to a fully-populated
506 * reservation and -1 otherwise.
509 vm_reserv_level_iffullpop(vm_page_t m)
513 rv = vm_reserv_from_page(m);
514 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
518 * Prepare for the reactivation of a cached page.
520 * First, suppose that the given page "m" was allocated individually, i.e., not
521 * as part of a reservation, and cached. Then, suppose a reservation
522 * containing "m" is allocated by the same object. Although "m" and the
523 * reservation belong to the same object, "m"'s pindex may not match the
526 * The free page queue must be locked.
529 vm_reserv_reactivate_page(vm_page_t m)
534 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
535 rv = vm_reserv_from_page(m);
536 if (rv->object == NULL)
538 KASSERT((m->flags & PG_CACHED) != 0,
539 ("vm_reserv_uncache_page: page %p is not cached", m));
540 if (m->object == rv->object &&
541 m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
542 vm_reserv_populate(rv);
544 KASSERT(rv->inpartpopq,
545 ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
547 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
548 rv->inpartpopq = FALSE;
549 LIST_REMOVE(rv, objq);
551 /* Don't vm_phys_free_pages(m, 0). */
552 m_index = m - rv->pages;
553 for (i = 0; i < m_index; i++) {
554 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
555 vm_phys_free_pages(&rv->pages[i], 0);
559 for (i++; i < VM_LEVEL_0_NPAGES; i++) {
560 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
561 vm_phys_free_pages(&rv->pages[i], 0);
565 KASSERT(rv->popcnt == 0,
566 ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
574 * Breaks the given partially-populated reservation, releasing its cached and
575 * free pages to the physical memory allocator.
577 * The free page queue lock must be held.
580 vm_reserv_reclaim(vm_reserv_t rv)
584 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
585 KASSERT(rv->inpartpopq,
586 ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", rv));
587 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
588 rv->inpartpopq = FALSE;
589 KASSERT(rv->object != NULL,
590 ("vm_reserv_reclaim: reserv %p is free", rv));
591 LIST_REMOVE(rv, objq);
593 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
594 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
595 vm_phys_free_pages(&rv->pages[i], 0);
599 KASSERT(rv->popcnt == 0,
600 ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", rv));
601 vm_reserv_reclaimed++;
605 * Breaks the reservation at the head of the partially-populated reservation
606 * queue, releasing its cached and free pages to the physical memory
607 * allocator. Returns TRUE if a reservation is broken and FALSE otherwise.
609 * The free page queue lock must be held.
612 vm_reserv_reclaim_inactive(void)
616 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
617 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
618 vm_reserv_reclaim(rv);
625 * Searches the partially-populated reservation queue for the least recently
626 * active reservation with unused pages, i.e., cached or free, that satisfy the
627 * given request for contiguous physical memory. If a satisfactory reservation
628 * is found, it is broken. Returns TRUE if a reservation is broken and FALSE
631 * The free page queue lock must be held.
634 vm_reserv_reclaim_contig(vm_paddr_t size, vm_paddr_t low, vm_paddr_t high,
635 unsigned long alignment, unsigned long boundary)
637 vm_paddr_t pa, pa_length;
641 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
642 if (size > VM_LEVEL_0_SIZE - PAGE_SIZE)
644 TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
645 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
646 if (pa + PAGE_SIZE - size < low) {
647 /* this entire reservation is too low; go to next */
651 for (i = 0; i < VM_LEVEL_0_NPAGES; i++)
652 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) {
653 pa_length += PAGE_SIZE;
654 if (pa_length == PAGE_SIZE) {
655 pa = VM_PAGE_TO_PHYS(&rv->pages[i]);
656 if (pa + size > high) {
657 /* skip to next reservation */
659 } else if (pa < low ||
660 (pa & (alignment - 1)) != 0 ||
661 ((pa ^ (pa + size - 1)) &
662 ~(boundary - 1)) != 0)
665 if (pa_length >= size) {
666 vm_reserv_reclaim(rv);
676 * Transfers the reservation underlying the given page to a new object.
678 * The object must be locked.
681 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
682 vm_pindex_t old_object_offset)
686 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
687 rv = vm_reserv_from_page(m);
688 if (rv->object == old_object) {
689 mtx_lock(&vm_page_queue_free_mtx);
690 if (rv->object == old_object) {
691 LIST_REMOVE(rv, objq);
692 LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
693 rv->object = new_object;
694 rv->pindex -= old_object_offset;
696 mtx_unlock(&vm_page_queue_free_mtx);
701 * Allocates the virtual and physical memory required by the reservation
702 * management system's data structures, in particular, the reservation array.
705 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
711 * Calculate the size (in bytes) of the reservation array. Round up
712 * from "high_water" because every small page is mapped to an element
713 * in the reservation array based on its physical address. Thus, the
714 * number of elements in the reservation array can be greater than the
715 * number of superpages.
717 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
720 * Allocate and map the physical memory for the reservation array. The
721 * next available virtual address is returned by reference.
723 new_end = end - round_page(size);
724 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
725 VM_PROT_READ | VM_PROT_WRITE);
726 bzero(vm_reserv_array, size);
729 * Return the next available physical address.
734 #endif /* VM_NRESERVLEVEL > 0 */